ATPO-LSTM: Adaptive Two-Phase Optimization with Entropy-Driven Genetic Algorithm for High-Dimensional LSTM Hyperparameter Tuning

Abstract

Aiming at the high-dimensional non-convex hyperparameter optimisation difficulties faced by long short-term memory networks (LSTMs) in time series forecasting, this paper proposes an adaptive two-stage optimisation framework based on entropy-driven genetic algorithm (ATPO-GA). The method dynamically switches the variational modes of Cauchy and Gaussian distributions through Shannon entropy, combines a hybrid selection strategy of tournament selection and simulated annealing, and introduces a dimension-aware adaptation assessment that includes computational complexity constraints, which effectively solves the problems of precocious convergence and diversity loss of the traditional evolutionary algorithms when dealing with more than 20-dimensional parameters. Experiments on 6 industrial datasets (energy, finance, healthcare, etc.) show that ATPO-LSTM reduces the mean absolute error (MAE) by 18.7% (p<0.01) and improves the convergence speed by 23% compared with particle swarm optimisation (PSO-LSTM). In a practical deployment in a regional grid system, 12.6% cost savings are achieved through accurate load forecasting. The theoretical analysis demonstrates the global convergence of the algorithm and maintains linear computational complexity in a 50-dimensional high-dimensional space. The results provide a new paradigm for efficient hyperparameter optimisation of industrial LSTM models.